Robot Bin-Picking Based On Domain Adaptation And Object Pose Estimation

It has become a novel solution in industrial automation to perform bin-picking with industrial robots and intelligent sensor.Accurate detection and object pose estimation from clutter scene is the key problem to solve bin-picking task for industrial robot.To handle the robot bin-picking problem in the clutter and complex scene,focusing on the piece picking task and the machine tending task,the methods of the 6D pose estimation base on transfer learning and grasp pose computation are researched in this paper.A robot bin-picking system using RealSense RGB-D sensor and UR5 robot is developed to verify the practical performance of our work.A pipeline for robot grasping based on transfer learning,6D pose estimation and grasp pose computation is proposed in this paper,which promotes the robot bin-picking in the clutter and complex scene without human annotations in the real world.Firstly,the labeled training samples in the simulation domain is generated.To handle the domain gap between the simulation domain and the real-world domain,a pixel-level domain adaption network named CCM Pixel-DA with cycle-content-mapping consistency is proposed without any real-world annotations and paired samples.Then,the CCM Pixel-DA is utilized for the Real-to-Sim transfer of the pose estimation RGB testing samples and the improved ClearGrasp is employed for the Real-to-Sim transfer of the pose estimation depth testing samples.The inference pipeline for 6D pose estimation based on transfer learning is proposed.In the grasp pose computation based on object pose,the grasp pose estimation method based on superquadric model is proposed for accurate and fast piece picking for mix objects.The grasp pose computation method based on registration is proposed for machine mending,which demands the special grasp pose.Based on the research above,the experiments of image transfer and 6D pose estimation on LineMOD dataset validate the good performance of the CCM Pixel-DA network and the inference pipeline.Furthermore,a robot bin-picking system including perception,interaction and planning is built in the real-world to verify the feasibility and availability of the proposed methods.